DreamVAR: Taming Reinforced Visual Autoregressive Model for High-Fidelity Subject-Driven Image Generation

TL;DR

The paper addresses the challenge of faithful subject-driven image generation by leveraging a Visual Autoregressive (VAR) framework with next-scale prediction. DreamVAR pre-fills multi-scale reference features from a visual tokenizer to mitigate train-test discrepancy inherent in autoregressive conditioning, and couples this with Group Relative Policy Optimization (GRPO) using dual rewards for subject consistency and semantic alignment. The authors implement a three-stage training protocol on Subject-200K and a high-quality DreamSubject-14K dataset, culminating in strong subject fidelity and competitive image-text alignment with only 2B parameters, as evidenced by superior DINO and CLIP-I scores and faster training/inference than larger diffusion baselines. This approach demonstrates that VAR with pre-filled conditioning and RL-based fine-tuning can achieve high-fidelity, subject-preserving image generation efficiently, with practical impact for real-time and scalable subject-driven synthesis.

Abstract

Recent advances in subject-driven image generation using diffusion models have attracted considerable attention for their remarkable capabilities in producing high-quality images. Nevertheless, the potential of Visual Autoregressive (VAR) models, despite their unified architecture and efficient inference, remains underexplored. In this work, we present DreamVAR, a novel framework for subject-driven image synthesis built upon a VAR model that employs next-scale prediction. Technically, multi-scale features of the reference subject are first extracted by a visual tokenizer. Instead of interleaving these conditional features with target image tokens across scales, our DreamVAR pre-fills the full subject feature sequence prior to predicting target image tokens. This design simplifies autoregressive dependencies and mitigates the train-test discrepancy in multi-scale conditioning scenario within the VAR paradigm. DreamVAR further incorporates reinforcement learning to jointly enhance semantic alignment and subject consistency. Extensive experiments demonstrate that DreamVAR achieves superior appearance preservation compared to leading diffusion-based methods.

Figures (3)

• Figure 1: (a) The framework of our DreamVAR for subject-driven image generation, which alleviates the train-test discrepancy through subject feature pre-filling. (b) The pipeline of reinforcement learning with GRPO in our DreamVAR.
• Figure 2: Qualitative comparisons with different methods on Dreambench.
• Figure 3: Ablation study of reinforcement learning rewards.